Royalty Driven LLM (RDLLM)

RDLLM adds visible source evidence and auditable creator allocation to AI answers. It records which sources were supplied, which claims they support, how much value was allocated, and whether settlement is eligible, held, or escrowed.

Try It

python -m pip install .
rdllm-first-run

The first line should be:

rdllm_first_run status: passed

The demo is deliberately synthetic. Its creators, works, URLs, revenue, and allocations are fictitious; no model provider is called and no money moves.

Follow the five-minute walkthrough

What You Get

• an answer with visible source labels such as [S1];
• claim-to-evidence rows with hashes and character spans;
• observable usage metrics for retrieval, matching, citations, and allocation;
• a source footer that survives API, copy, and export verification;
• a settlement decision that fails closed when evidence is weak or missing;
• publicly verifiable Ed25519 receipts for production integrations.

Example footer row:

[S1] Provenance Ledgers for AI Outputs; claims=2; support=0.900;
weight=1.000000; payout=0.550000; settlement=allocated_not_executed

payout is an allocation amount, not proof that a payment executed. Direct settlement additionally requires a trusted external payment-processor attestation.

Three Ways To Use It

Ground a provider response: RDLLM retrieves allowed source context before generation, requires the provider to return matching source IDs or native annotations, and binds request, response, context, and evidence hashes.

Audit existing text: RDLLM can measure observable overlap after generation, but labels that result post_hoc_observable_match and holds settlement for review. It does not present similarity as proof of model reliance.

Allocate creator value: verified context-grounded claims can produce a settlement instruction. Unsupported, negated, numerically embellished, disputed, or rights-blocked claims route value to escrow.

See 15 runnable use cases with screenshots

Trust Boundary

RDLLM 1.0 is production-stable software. An operator deployment is not production-attested merely because its configuration passes. Production claims require signed third-party evidence verified against an independently managed trust store. Bundled artifacts are synthetic fixtures and shared-secret HMAC is never accepted as a public production signature.

RDLLM reports observable source support and allocation by default:

observable_support_allocation_not_model_internal_reliance

It does not claim to discover hidden training-data use or causal model reliance from output similarity.

Integrate

• API guide
• API client examples
• Provider onboarding
• Deployment and external attestations
• Receipt protocol

Understand

• GitHub Start Here
• Quickstarts: English, French, Vietnamese
• Plain-language and multilingual explanations
• Mechanism and formulas
• Evidence and limitations
• Recent research
• White paper
• Project attribution map

Verify The Release

PYTHONPATH=src python3 -m unittest discover -s tests
PYTHONPATH=src python3 tools/build_public_site.py --check
PYTHONPATH=src python3 tools/production_readiness.py --json
PYTHONPATH=src python3 tools/ship_check.py

The repository release may be ready while production_grade_claim_allowed is false. That is expected for the bundled unattested profile and prevents the software from certifying its own deployment.

Version 1.0.0. MIT licensed. Created and maintained by Siddharth Nilesh Patel.